|Publication number||US7260682 B2|
|Application number||US 11/188,668|
|Publication date||Aug 21, 2007|
|Filing date||Jul 25, 2005|
|Priority date||Jul 27, 2004|
|Also published as||EP1622009A1, US7493476, US7500085, US7533250, US7543285, US7546437, US7574584, US7587583, US7606977, US7624382, US7743384, US7752610, US7757223, US7930689, US8024554, US8024716, US8046748, US8078842, US8185666, US8380906, US8516496, US9201807, US20060023517, US20060025986, US20060026126, US20060026183, US20060026200, US20060026201, US20060026312, US20060026322, US20060026353, US20060026354, US20060026357, US20060026370, US20060026390, US20060026391, US20060026392, US20060026393, US20060026394, US20060026395, US20060026396, US20060026397, US20060026398, US20060026400, US20060026401, US20060026402, US20060026403, US20060026404, US20060026405, US20060026407, US20060026412, US20060026563, US20060026564, US20060026565, US20060026566, US20060026571, US20060026574, US20060026575, US20060026580|
|Publication number||11188668, 188668, US 7260682 B2, US 7260682B2, US-B2-7260682, US7260682 B2, US7260682B2|
|Inventors||Jean-Philippe Lesot, Gilbert Cabillic, Gerard Chauvel|
|Original Assignee||Texas Instruments Incorporated|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (5), Referenced by (1), Classifications (15), Legal Events (3)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application claims priority to European Patent Application No. 04291918.3, filed on Jul. 27, 2004 and incorporated herein by reference. This application also contains subject matter that may be related to U.S. patent applications Ser. No. 10/818,584 entitled “Management of Stack-Based Memory Usage in a Processor, Ser. No. 10/632,067 entitled “Memory Management of Local Variables,” Ser. No. 10/632,076 entitled “Memory Management of Local Variables Upon a Change of Context,” Ser. No. 10/632,228 entitled “System and Method to Automatically Stack and Unstack Java Local Variables.” This applications also contains subject matter that may be related to concurrently filed applications entitled “Memory Usable in Cache Mode or Scratch Pad Mode to Reduce the Frequency of Memory Accesses” and “Context Save and Restore With a Stack-Based Memory Structure”.
1. Technical Field of the Invention
The present disclosure relates generally to processors and more particularly to the use of cache memory as scratch pad storage.
2. Background Information
Many types of electronic devices are battery operated and thus preferably consume as little power as possible. An example is a cellular telephone. Further, it may be desirable to implement various types of multimedia functionality in an electronic device such as a cell phone. Examples of multimedia functionality may include, without limitation, games, audio decoders, digital cameras, etc. It is thus desirable to implement such functionality in an electronic device in a way that, all else being equal, is fast, consumes as little power as possible and requires as little memory as possible. Improvements in this area are desirable.
In at least some embodiments, a processor adapted to couple to external memory. The processor comprises a controller and data storage. The data storage is usable to store local variables and temporary data and is configurable to operate in either a cache policy mode in which a miss results in an access of the external memory or in a scratch pad policy mode in which a miss does not result in an access of the external memory. The data storage comprises first and second portions, and wherein only one of said portions is active at a time for storing said local variables. When the active portion does not have sufficient capacity for additional local variables, the other portion becomes the active portion for storing local variables. When one portion is the active portion, the other portion is used to store the temporary data and such other portion is sufficiently large to contain the temporary data. Other embodiments comprise a system (e.g., a cellular telephone) containing such a processor and an associated method.
Certain terms are used throughout the following description and claims to refer to particular system components. As one skilled in the art will appreciate, different companies may refer to a component by different names. This document does not intend to distinguish between components that differ in name but not function. In the following discussion and in the claims, the terms “including” and “comprising” are used in an open-ended fashion, and thus should be interpreted to mean “including, but not limited to . . . ”. Also, the term “couple” or “couples” is intended to mean either an indirect or direct connection. Thus, if a first device couples to a second device, that connection may be through a direct connection, or through an indirect connection via other devices and connections. The terms “first portion” and “second portion” are intended to broadly refer to either portion of the multi-portion RAMset explained below.
For a more detailed description of the preferred embodiments of the present invention, reference will now be made to the accompanying drawings, wherein:
The following discussion is directed to various embodiments of the invention. Although one or more of these embodiments may be preferred, the embodiments disclosed should not be interpreted, or otherwise used, as limiting the scope of the disclosure, including the claims, unless otherwise specified. In addition, one skilled in the art will understand that the following description has broad application, and the discussion of any embodiment is meant only to be exemplary of that embodiment, and not intended to intimate that the scope of the disclosure, including the claims, is limited to that embodiment.
The subject matter disclosed herein is directed to a programmable electronic device such as a processor having memory in which “local variables” associated with a stack-based language (e.g., Java) and pointers associated with the local variables may be stored. The term “local variables” refers to temporary variables used by a method that executes on the processor. Multiple methods may run on the processor and each method preferably has its own set of local variables. In general, local variables have meaning only while their associated method is running. The stack-based language may comprise Java Bytecodes although this disclosure is not so limited. In Java Bytecodes, the notion of local variables (“LVs”) is equivalent to automatic variables in other programming languages (e.g., “C”) and other termed variables in still other programming languages. This disclosure, however, is not limited to Java, Java methods, and Java local variables. The principles disclosed below are applicable to any system that manages a stack and includes “put block” and “pop block” operations to push a block of data onto a stack or pop a block of data from a stack.
The following describes the operation of a preferred embodiment of such a processor in which the methods and local variables may run and be used. Other processor architectures and embodiments may be used and thus this disclosure and the claims which follow are not limited to any particular type of processor.
The processor described herein is particularly suited for executing Java™ Bytecodes, or comparable code. As is well known, Java is particularly suited for embedded applications. Java is a relatively “dense” language meaning that on average each instruction may perform a large number of functions compared to various other programming languages. The dense nature of Java is of particular benefit for portable, battery-operated devices that preferably include as little memory as possible to save space and power. The reason, however, for executing Java code is not material to this disclosure or the claims that follow.
Referring now to
Referring again to
The JVM 108 generally comprises a combination of software and hardware. The software may include the compiler 110 and the hardware may include the JSM 102. The JVM may include a class loader, bytecode verifier, garbage collector, and a bytecode interpreter loop to interpret the bytecodes that are not executed on the JSM processor 102.
Referring now to
Referring again to
The data storage 122 generally comprises data cache (“D-cache”) 124 and a data random access memory (“D-RAMset”) 126. The D-RAMset (or simply “RAMset”) 126 preferably comprises one “way” of the multi-way cache. Reference may be made to co-pending applications U.S. Ser. No. 09/591,537 filed Jun. 9, 2000 Ser. No. 09/591,656 filed Jun. 9, 2000 and Ser. No. 09/932,794 filed Aug. 17, 2001, all of which are incorporated herein by reference. The stack (excluding the micro-stack 146), arrays and non-critical data may be stored in the D-cache 124, while Java local variables and associated pointers as explained below, as well as critical data and non-Java variables (e.g., C, C++) may be stored in D-RAMset 126. The instruction storage 130 may comprise instruction RAM (“I-RAMset”) 132 and instruction cache (“I-cache”) 134. The I-RAMset 132 may be used to store “complex” micro-sequenced Bytecodes or micro-sequences or predetermined sequences of code.
In accordance with a preferred embodiment of the invention, at least some applications executed by the JSM 102 comprise one or more methods. A “method” includes executable instructions and performs one or more functions. Other terms for “method” may include subroutines, code segments, and functions, and the term should not be used to narrow the scope of this disclosure.
A method (the “calling” method) may call another method (the “called” method). Once the called method performs its function, program control returns to the calling method. Multiple hierarchical levels of methods are possible as illustrated in
A method may have one or more “local variables,” as explained previously. Local variables may be used to temporarily store data or other information as the method performs its task(s). The local variables preferably are specific to the method to which the variables pertain. That is, method A's local variables (“LVA”) are accessible generally by only method A and have meaning only to method A. Once method A completes, the method A local variables become meaningless. Similarly, LVB and LVC comprise local variables associated with methods B and C, respectively. Java Bytecodes refer to local variables using an index. The JVM maintains a local variables pointer (“PTR LV”) which points to the base address of the memory containing the current method's local variables. To access a particular local variable, a suitable index value is added to the base address to obtain the address of the desired local variable. In general, the local variables associated with one method may have a different size than the local variables associated with another method.
Although a plurality of methods may run on the JSM 102, typically only one method is “active” at a time having its instructions actively being executed by the JSM 102. The base pointer of the currently active method preferably is stored in register R5 as noted previously. In general, the base pointer for the active method may be computed by the JVM 108 while executing the invoke bytecode of the active method.
Sequence 510 depicts the state of the D-RAMset 126 when method A calls method B. In accordance with the preferred embodiments of the invention, the local variables (LVB) associated with method B are stacked in storage space 512 generally adjacent LVA (“on top of” LVA when viewed as in
Following arrow 507 to time sequence 520, when method C is invoked (called by method B), the base pointer for method B (PTR LVB) is stored in location 514A which may be on top of LVB and below PTR LVC as shown and register R5 is loaded with the base pointer 524 (PTR LVC) to the base of the LVC data set 522. Method C's local variables (LVC) are allocated to storage space 522 which generally is adjacent (on top of) LVB 512 and PTR LVB 514A as shown. The PTR LVB value is stored in location 514A according to a similar calculation as that described above.
In accordance with preferred embodiments of the invention, the D-RAMset 126 is configured to provide any one or more or all of the following properties. The implementation of the D-RAMset 126 to provide these properties is explained in detail below. The local variables and pointers stored in the D-RAMset 126 preferably are “locked” in place meaning that, although the D-RAMset 126 is implemented as cache memory, eviction of the local variables generally can be prevented in a controlled manner. The locking nature of the D-RAMset 126 may be beneficial while a method executes to ensure that no cache miss penalty is incurred. Additionally, write back of valid, dirty local variables to main memory 106 is avoided in at least some situations (specified below). Further, mechanisms can be employed in the event that the D-RAMset 126 has insufficient capacity to accommodate all desired local variables. Further still, once a method has completed, the portion of the D-RAMset allocated for the completed method's local variables remains marked as “valid.” In this way, if and when such methods or any new methods are executed and re-use the RAMset space (such as that described in one or more of the copending applications mentioned above), such methods' associated local variables will be mapped to the same portion of the D-RAMset. If the RAMset lines are already marked as valid, access to those new local variables may not generate any misses. Retrieval of data from memory in this situation is unnecessary because the local variables only have significance while a method executes and a newly executing method first initializes all of its local variables before using them. Not generating misses and thus avoiding fetching lines from external memory reduces latency and power consumption. After a relatively short period of time following the start of a Java program execution, all relevant lines of the RAMset are marked as valid and accesses to local variables of newly called methods do not generate misses, thereby providing superior performance of a “0-wait state memory.” Furthermore, the cache properties of RAMset allow discarding or saving of the data in main memory whenever required.
In accordance with a preferred embodiment of the invention, the local variables (LVA-LVC) and associated pointers (PTR LVA-PTR LVC) may be stored in D-RAMset 126. The D-RAMset 126 may be implemented in accordance with the preferred embodiment described below and in copending applications entitled “Cache with multiple fill modes,” filed Jun. 9, 2000, Ser. No. 09/591,656; “Smart cache,” filed Jun. 9, 2000, Ser. No. 09/591,537; and publication no. 2002/0065990, all of which are incorporated herein by reference.
As described in greater detail below, in the preferred embodiment, the data storage 122 (
In operation, the processor's core 102 may access main memory 106 (
As shown, cache controller 222 couples to, or otherwise acceses, Full_Set_Tag registers 232 (individually referenced as registers 232 a through 232 c), Global_Valid bits 234 (individually referenced as bits 234 a through 234 c), tag memories 236 (individually referenced as tag memories 236 b and 236 c), valid entry bit arrays 237 (individually referenced as bit arrays 237 a through 237 c) and data arrays 238 (individually referenced as data arrays 238 a through 238 c). Comparators 240 (individually referenced as comparators 240 a through 240 c) may couple to respective Full_Set_Tag registers 232. Comparators 242 (individually referenced as comparators 242 b and 242 c) couple to respective tag memories 236. Output buffers 244 (individually referenced as buffers 244 a through 244 c) may couple to respective data arrays 238. HiVMiss logic 246 (individually referenced as logic 246 a through 246 c) may couple to comparators 240, global valid bits 234, valid bits 237, RAM_fill_mode bit 224 and Cache_Enable bit 226.
In operation, data storage 122 may be configured using the control bits 224, 226, 228 and 230. The Cache_Enable 226 allows the data storage to be enabled or disabled, as in standard cache architecture. If the data storage 122 is disabled (e.g., Cache Enable=0), data read accesses may be performed on the main memory 106 without using the data storage 122. If the data storage 122 is enabled (e.g., Cache_Enable=1), data may be accessed in the data storage 122, in cases where such data is present in the data storage. If a miss occurs, a line (e.g., 16 bytes) may be fetched from main memory 106 and provided to the core 120.
The size of the data array 238 a may be different than the size of the data arrays 238 b, c for the other ways of the cache. For illustration purposes and without limiting this disclosure in any way, it will be assumed that data arrays 238 b and 238 c are each 8 Kbytes in size, configured as 512 lines, with each line holding eight two-byte data values. Data array 238 a may be 16 Kbytes in size, configured as 1024 lines, each line holding eight, two byte data values. The ADDR[L] signals may be used to address one line of the data array 238 and valid bit array 237 (and tag memory 236, where applicable). Accordingly, for the 1024-line first way, ADDR[L] may include 10 bits [13:4] of an address from the core. For the 512-line second and third ways, ADDR[L] may include 9 bits [12:4] of an address from the core. The ADDR[H] signals define which set is mapped to a line. Thus, assuming a 4 Gbyte address space, ADDR[H] uses bits [31:14] of an address from the core for the first way and uses bits [31:13] for each of the second and third ways of the cache 130.
The tag memories 236 and comparators 242 may be used for a two-way set associative cache (e.g., D-cache 124 in
The RAMset cache 126 preferably stores data associated with a contiguous block of main memory 106 starting at an address defined by the Full_set_tag register 232 for the RAMset. This contiguous block of information (e.g., local variables/pointers) may be mapped to the corresponding data array 238 of the RAMset. In at least some embodiments, only the high order bits of the starting address are stored in the Full_set_tag register 232.
Referring again to
A RAMset hit situation may occur when the high order bits of the address from the core 120 match the contents of the Full_set_TAG register 232 and the global valid bit equals “1” (the setting of the global valid bit is described in greater detail below). By default, the RAMset comparison preferably has higher priority than the other cache ways. A hit situation indicates that the requested data is mapped into the RAMset. If the Valid entry bit 237 corresponding to the line containing the data is set to “1”, comparator 240 causes hit/miss logic 246 to generate a “hit-hit” signal because the address hit the RAMset and the data is present in the RAMset. If the corresponding valid bit 237 of the RAMset entry is “0”, logic 240 generates a “hit-miss” because the address hit the RAM set, but the data is not yet present in the RAM set. In this latter case, the data may be fetched from main memory 106 and loaded into the data array 238 of the RAMset. A hit in the RAMset logic preferably takes precedence over the normal cache logic. The standard logic of the two-way cache generates a miss when the RAMset logic generates a hit. Information can reside in both the RAMset and the two-way cache without causing any misbehavior; the duplicated cache entry in the 2-way cache will eventually be evicted by the replacement mechanism of the two-way cache because such data will not be used. However, in the preferred embodiment the data mapped onto a RAMset is first removed from the cache to avoid a data coherency problem. When configured as a RAMset, data array 238 a, b, c can be configured as a local RAM or as a cached segment depending on the setting of a suitable configuration bit (e.g., LR/C bit 231). However, even when configured as a local RAM, individual valid bits may be updated but misses do not generate accesses to the external memory.
To configure a RAMset for operation, the Full_set_tag register 232 preferably is loaded with a start address (set_start_addr) and the RAM_fill_mode bit 224 is configured to a desired fill mode. The circuitry for filling the cache can be the same as that used to fill lines of the set associative cache. At least one fill mode may be implemented and is referred to as a “line-by-line” fill mode as described below. Other fill modes may be implemented if desired such as the “set fill” mode described in at least one of the documents incorporated by reference.
For the line-by-line fill (RAM_fill_mode=0), the global valid bit 34 is set to “1” and each of the valid entry bits 237 is set to “0” when the Full_set_tag register 232 is loaded with the starting address. At this point, the data array 238 is empty (it is assumed that the Cache_Enable bit 226 is set to “1” to allow operation of the data storage 122). Upon receiving an address from the core 120, a valid entry bit 237 is selected based on the low order bits of the address. As provided above, if the RAMset is 16 Kbytes in size, organized as an array of 1 K×16 bytes, where 16 bytes is equivalent to a block line in the associated 2-way cache, the Full_set_TAG register 232 may store 18 bits [31:14] of the starting address. The address indexing each entry of the RAMset (ADDR[L]) may include 10 bits [13:4] while the data address used to access one data value in the line may include 4 bits [3:0] (assuming data accesses are 1 byte). In Java, local variables comprise four byte entities but, as explained previously, the RAMset may be shared between local variables and other, possibly critical, data. A line of the data array 238 (at ADDR[L]) is loaded from main memory 106 each time that a miss situation occurs because the comparator 240 determines a match between ADDR[H] and the content of Full_set_TAG, the Global valid bit 34 is set to “1” and the valid bit 237 associated with the line at ADDR[L] is “0”. The state of the RAMset in this mode of operation is also referred to as the cache policy “CP” state. This situation indicates that the selected line is mapped to the RAMset, but has not yet been loaded into the RAMset's data array 238. When the line is loaded into the data array 238 from main memory 106, the valid bit 237 corresponding to the line is set to “1”.
This loading procedure (resulting in the valid bit being set to indicate the presence of valid data) has the same time penalty as a normal cache line load, but the entry will remain locked in the RAMset (i.e., the valid bit will remain set) unless the content of the Full_Set_Tag is changed and, therefore, the processing device will not be penalized on a subsequent access. As such, the lines used by a completed method remain valid so that re-using the lines by subsequent methods does not necessitate accesses to main memory 106. Further, freeing the local variable space for a completed method generally only involves disregarding the relevant base pointer. Further still, there is no need to copy back local variables upon to main memory 106 upon completion of a method because such extinct local variables are not used any more.
In some situations, the capacity of the D-RAMset 126 may not be sufficient to hold all desired local variables. In accordance with at least one embodiment, excess local variables may be stored in the non-D-RAMset data arrays 238. In accordance with other embodiments, a larger block of local variables (i.e., larger than just the excess local variables) may be mapped to the non-D-RAMset cache ways. During the “invoke” bytecodes, that initiates a method call, the local variable size of the called method is known by the JVM 108. The JVM also knows the total RAMset size (via a readable configuration register) and the RAMset size already utilized. Therefore, based on this information, the JVM may or may not decide to map the new local variable area onto the RAMset. A method may have a large chunk of local variables and not use them on each call. Therefore, mapping those local variables onto the RAMset may force unnecessary RAMset management of the base pointer and saving/restoring of local variables of calling methods or may cause more frequent overflow of a subsequently called method. Instead, the JVM 108 may map the methods with larger chunks of local variables onto the non-RAMset data cache and thus preserve more space in the RAMset for methods with a smaller number of local variables. In some embodiments, many methods may have less than 10 local variables and almost all methods have less than about 40 local variables, but, of course, these numerical characterizations are application dependent. For methods with many local variables, the system may map those local variables outside the RAMset avoiding penalizing other methods. This technique is generally transparent for the return mechanism because of the management of the PTR_LV of the calling method. Upon completion of a method, the lines containing that method's local variables may remain marked as valid. As noted above, maintaining such lines marked as valid avoids generating misses in calls of new methods.
In accordance with some embodiments, more than one contiguous block of external memory 106 may be mapped onto the D-RAMset's data array 238. As illustrated in
A plurality of commands may be implemented in connection with the data storage 122. Such commands may include, without limitation, D-RAMset-Clean, D-RAMset-Flush, and D-RAMset-policy-set. In addition to valid bits 237 for each line, a dirty bit also may be provided to indicate whether or not the line contains dirty data. The D-RAMset-Clean command may be performed by examining the valid and dirty bits associated with each line. The D-RAMset-Clean command then copies back to external memory 106 only those lines that have valid and dirty data. In embodiments without dirty bits, the D_RAMset-Clean preferably copies all valid entries from D_RAMset 126 to external memory 106. The D-RAMset-Flush command invalidates lines within the D-RAMset 126 by clearing the relevant valid bits 237. The D-RAMset-Clean and D-RAMset-Flush commands may be performed in one of at least three variations. In one-variation, the D-RAMset-Clean and D-RAMset-Flush commands perform their respective actions on all of the lines of the D-RAMset 126 (D-RAMset-CleanAll and D-RAMset-FlushAll). In another variation, the D-RAMset-Clean and D-RAMset-Flush commands perform their respective actions on just those lines in the D-RAMset 126 that fall within a range of addresses specified as operands in the commands (D-RAMset-CleanRange and D-RAMset-FlushRange). A third variation permits the D-RAMset-Clean and D-RAMset-Flush commands to act on a single address within the D-RAMset 126 (D-RAMset-CleanEntry and D-RAMset-FlushEntry) providing the corresponding data address to be saved or invalidated.
One or more commands are used to specify whether a data array 238 configured as a RAMset is to function as a Local RAM or as cache. The D-RAMset-policy-set command is used in this regard. In the embodiments described below, this command is implemented as two separate commands called the SPP command and the CP command. Such commands may set one or bits in a register to indicate how a data array 238 is to be used. The bit that is set may comprise a bit in the status register R15, the LR/C bit 231 in a register in the cache controller 222 (
Referring now to
With the structure described above, less than desirable behavior can occur in a particular situation. The situation is when the RAMset is full and a clean operation is performed to write its data to the associated page of memory to make room for additional local variables in the RAMset. The cleaning process takes time and consumes power. A new method is invoked and then uses the newly mapped RAMset for its local variables. Returning from this method to the prior method entails flushing the RAMset and bringing the previously saved local variable data back into the RAMset from memory. This flushing and retrieval from memory process also takes time and consumes power. There can be situations in which the RAMset is cleaned to make room for new data, the RAMset is used for the new data, but returns back to the prior set local variables (saved to memory) relatively quickly. In fact, one can imagine a loop in the executable code in which a method invokes a new method each time through the loop. This repeated invocation of the new method may entail a clean operation and the exit from the new method back to the calling method will a corresponding flush and memory retrieval of the calling method's data. This repeated invocation of a called method and return back to the calling method (an “oscillation”) at the boundary of the RAMset space (thereby forcing a clean, flush, etc.) can consume considerable power and time just cleaning the RAMset and then flushing and bringing the data back into the RAMset. The following embodiment solves this problem.
In accordance with a preferred embodiment,
The embodiment of the RAMset in multiple portions uses the commands listed in Table I.
Switch RAMset to scratch pad policy
Switch RAMset to cache policy
Clean upper portion of RAMset to memory
Clean lower portion of RAMset to memory
Invalidate upper portion
Invalidate lower portion
Allocate new memory page and set RAMset
base address accordingly in Full_Set_Tag
Free current memory page and restore RAMset
base address to previous base address in
The SPP and CP commands cause the RAMset to be in the SPP and CP modes as discussed previously. The UPPER CLEAN and LOWER CLEAN commands can be implemented using the D-RAMset-CleanRange and D-RAMset-CleanEntry commands to clean just the upper or lower portions, respectively. Similarly, the UPPER FLUSH and LOWER FLUSH commands can be implemented using the D-RAMset-FlushRange and D-RAMset-FlushEntry commands to flush just the upper or lower portions, respectively. The R.SET(++) command causes a new page of external memory 106 to be allocated and mapped to the RAMset using the base address of the new memory page. The previous base address of the RAMset is saved as part of the data in the RAMset. The R.SET(−−) command essentially performs the reverse operation of the R.SET(++) command and frees the current external memory page while restoring the base address of the RAMset to the previous base address.
The RAMset may initialize into state 700. In state 700, the RAMset is in the SPP mode to permit the upper portion to be used to store data (e.g. local variables) but to avoid accesses to external memory 106 upon a cache miss. As explained above, a JAVA method typically requires an allocation of a portion of the RAMset for use for its local variables. Further, one method may invoke another method which, in turn, may invoke another method, and so on. Each such invoked method requires a new allocation of storage space in the RAMset. In state 700, each such allocation falls within the upper portion which is the active portion.
At some point, however, an invocation of a new method may require an allocation of RAMset storage that may exceed the available unused capacity of the upper portion. At this point, the lower portion of the RAMset needs to be used to store additional local variables for the newly invoked method. The invocation of this new method is identified by arrow 701 which points to RAMset state 702.
In RAMset state 702 (which is also in operated in the SPP mode), the lower portion of the RAMset is now the active portion. The lower portion therefore can be used to store local variables for the newly invoked method and any additional methods that are invoked therefrom. As explained above, each called method returns to its calling method. As such, the method that was invoked that caused the transition from the upper portion being active to the lower portion of the RAMset being active may eventually return to the calling method. The return to such method is illustrated with arrow 703. Further, an oscillation may occur between such methods—the method that invoked a method causing the transition to the lower portion as well as the transition back from such method. This type of oscillation (identified by oppositely pointing arrows 701 and 703 in dashed circle 690), however, is not as problematic as the oscillations noted above because the oscillation identified by arrows 701 and 703 do not require cleaning, flushing, or re-loading the RAMset. That is, no memory access is required to oscillate between the two RAMset states 700 and 702. Because no memory accesses are required, such oscillations advantageously take less time and consume less power.
However, as more and more methods are invoked requiring allocations of the lower portion of the RAMset while in state 702, eventually, the entire RAMset (i.e. both portions) may become full of valid data. At this point, any new method that is invoked will require an allocation of RAMset space greater than the available space to be allocated in the RAMset. Consequently, a portion of the RAMset is cleaned (i.e. copied to external memory 106) to make room for new data. This cleaning process is illustrated by arrow 705 which points to RAMset state 704. In particular, the clean operation only cleans the upper portion of the RAMset. The data in the upper portion represents the oldest data in the RAMset and is copied to the corresponding page of external memory. The R.set(++) command is also performed at this time to allocate a new external memory page to the RAMset.
At state 704 the upper portion of the RAMset can again be used to store new local variables for newly invoked methods. The upper portion therefore becomes the active portion of the RAMset. At this point, the upper portion of the RAMset is the active portion, the lower portion of the RAMset contains valid data but is not currently used as the active portion, and the initial data in the upper portion from state 700 (or other data from states 706 or 710) has been copied to external memory.
If insufficient space in the upper portion is available for the local variables of additional methods to be invoked, the lower portion of the RAMset can then be used for such additional local variables. In state 704, however, the lower portion of the RAMset may already have valid local variables and thus a clean operation (II.CLEAN command) is performed to first clean the lower portion so that the lower portion can be used for additional local variables. This process is depicted via arrow 707 which points back to state 702.
While at state 704 (also in SPP mode), new methods can be invoked and allocations of storage space in the upper portion of the RAMset can be performed for usage by such new methods. Of course, called methods may return back to their previous calling methods and eventually, the method that caused the first allocation of the upper portion at state 704 may return back to its calling method. That return is illustrated by arrow 711, which points to RAMset state 706 (also in SPP mode). At RAMset state 706, therefore, the lower portion of the RAMset again becomes the active portion. From the lower portion in state 706, a method may be invoked which again exceeds the available capacity of the lower portion thereby causing the upper portion to become the active portion as identified by arrow 709 which transitions back to state 704. Again, an oscillation can occur between states 704 and 706 (identified by oppositely pointing arrows 709 and 711 in dashed circle 691), but such oscillations do not require any memory accesses and therefore can be performed with little time and little power consumption.
From state 706, with the bottom portion being active, if a return is to be performed to a prior method whose local variables were stored in the upper portion of the RAMset and such data has been copied to external memory 106 (in a prior clean operation of the upper portion), the RAMset transitions to state 708 by way of return arrow 713. Because the data associated with upper portion of the RAMset has been saved off to external memory, a flush of the upper portion is performed to invalidate the upper portion. Further, the upper portion of the RAMset, now the active is transitioned to the CP mode to permit the previously saved data to be loaded into the RAMset's upper portion.
From state 708, if a return is performed to a prior method whose local variables are associated with the lower portion of the RAMset but have been saved off to external memory, the RAMset operates according to state 714 still in the CP mode (arrow 721). A R.SET(−−) command is performed to free the current memory page and restore the RAMset base address to the previous base address. Also, a flush of the bottom portion if performed to cause the bottom portion's data to be retrieved from external memory.
Going back to state 708, if RAMset storage space is needed for a new method and the extra storage is not available in the currently active upper portion, the RAMset operates according to state 710. In state 710, the RAMset operates in the SPP mode and the bottom portion becomes the active portion for storing local variables. This invocation is illustrated by arrow 715. A return to the method that caused the bottom portion to become active may be performed back to state 708 (arrow 717). An oscillation between states 708 and 710, designated by oppositely pointing arrows 715, 717 within dashed circle 693 do not require any external memory accesses and therefore can be performed in relatively little time and with relatively little power consumption.
From state 710, an invocation of a method that exceeds the storage capacity of the active lower portion takes the RAMset to a different state, in particular, state 704. This transition is shown by way of arrow 719 and also requires a clean of the upper portion to be performed to save the data already present in the upper portion so that the upper portion of the RAMset can be used for additional local variables.
From state 700, a return to a method whose local variables are associated with the lower portion but have been saved to external memory can be performed with the RAMset now operating to state 714. This transition is identified by arrow 729 and a flush of the lower portion is performed along with a change in the allocation policy to the CP mode. The change to the CP mode causes previously cleaned data from external memory to be re-loaded into the corresponding lines of the lower portion of the RAMset.
An oscillation can also occur between states 712 and 714 between the lower and upper portions of the RAMset. The oscillations are indicated by oppositely pointing invocation arrow 725 and return arrow 727 within dashed circle 694. This oscillation occurs without accesses to external memory and thus requires little time and power. As with the oscillation between states 708 and 710, the oscillation between states 712 and 714 require a change in allocation policy as shown. RAMset state 712 is in the SPP mode because the needed local variable data is already in the upper portion. State 714 is in the CP mode because the needed data must be retrieved from external memory and re-loaded into the lower portion of the RAMset.
From state 712, an invocation of a method that exceeds the storage capacity of the active upper portion takes the RAMset to a different state, and in particular, state 702. This transition is shown by way of arrow 731 and also requires a clean of the lower portion to be performed to save the data already present in the lower portion so that the lower portion of the RAMset can be used for additional local variables.
Finally, from state 714 in which the lower portion is active, a method that returns to a calling method whose local variable data is stored in the upper portion causes state transition to state 708 (arrow 723). This transition makes the upper portion the active portion so that the upper portion can be used to access the local variables stored therein.
In accordance with at least one embodiment of the invention, a state variable is maintained to indicate the state of the RAMset. For example,
In some situations, it may be desirable to have access to fast memory storage for storing temporary data. Such data may be used, for example, in multi-media computations. Such data need not necessarily comprise Java local variables as described above. The following embodiments describe the use of at least a portion of the RAMset for storing temporary, transient data. For these embodiments, it is assumed that a thread switch will not occur and that the portion of the RAMset being used for such data is sufficiently large for the data to be stored therein. In other words, it will not be necessary to clean a portion of the RAMset to make room for additional such temporary data when the portion currently being used proves too small to accommodate all needed temporary data—the portion of the RAMset used for this temporary data is big enough to avoid this problem.
While the RAMset states on the right-hand side of
While the preferred embodiments of the present invention have been shown and described, modifications thereof can be made by one skilled in the art without departing from the spirit and teachings of the invention. The scope of protection is not limited by the description set out above. Each and every claim is incorporated into the specification as an embodiment of the present invention.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5276835 *||Dec 14, 1990||Jan 4, 1994||International Business Machines Corporation||Non-blocking serialization for caching data in a shared cache|
|US5537574 *||Mar 30, 1992||Jul 16, 1996||International Business Machines Corporation||Sysplex shared data coherency method|
|US5983313 *||Apr 10, 1996||Nov 9, 1999||Ramtron International Corporation||EDRAM having a dynamically-sized cache memory and associated method|
|US6510493 *||Jul 15, 1999||Jan 21, 2003||International Business Machines Corporation||Method and apparatus for managing cache line replacement within a computer system|
|US6574708 *||May 18, 2001||Jun 3, 2003||Broadcom Corporation||Source controlled cache allocation|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US20110131381 *||Nov 27, 2009||Jun 2, 2011||Advanced Micro Devices, Inc.||Cache scratch-pad and method therefor|
|U.S. Classification||711/118, 711/154, 711/100, 711/E12.067, 711/170|
|International Classification||G06F12/00, G06F13/00|
|Cooperative Classification||G06F9/45504, G06F12/0802, Y02B60/1225, G06F9/30174, G06F2212/6012, G06F12/1081|
|European Classification||G06F9/30U2, G06F12/10P|
|Jul 25, 2005||AS||Assignment|
Owner name: TEXAS INSTRUMENTS INCORPORATED, TEXAS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:LESOT, JEAN-PHILIPPE;CABILLIC, GILBERT;CHAUVEL, GERARD;REEL/FRAME:016812/0461;SIGNING DATES FROM 20050718 TO 20050720
|Jan 3, 2011||FPAY||Fee payment|
Year of fee payment: 4
|Dec 31, 2014||FPAY||Fee payment|
Year of fee payment: 8